---
title: What requirements should an enterprise set for managing AI agent skills?
url: https://opencontroller.lyzr.com/blog/enterprise-requirements-for-ai-agent-skills-management
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# What requirements should an enterprise set for managing AI agent skills?

**Write skills requirements across six lifecycle areas: inventory, authoring standards, review and approval, distribution and access, runtime monitoring, and change and retirement. Each requirement should have an acceptance test you can run in a demo or pilot. Without tests, a requirements list becomes a checkbox exercise that every vendor passes.**

Skills are packaged instructions, scripts and resources that an agent loads when a task needs them. Anthropic's Agent Skills format centres on a `SKILL.md` file with YAML frontmatter. The required fields are `name` (lowercase letters, numbers and hyphens, up to 64 characters) and `description` (up to 1,024 characters). Skills now run in chat apps, in coding agents on laptops and in agents built with SDKs. Enterprises that took them seriously have written long requirement lists, some with dozens of items. The list below groups the requirements that come up most and adds a test for each.

## 1. Inventory: do you know every skill that exists?

| Requirement | Acceptance test |
|---|---|
| Discover skills provisioned at organisation level | Provision a test skill; it appears in the inventory |
| Discover user-created and shared skills | A user uploads a skill; it appears with the user as creator |
| Discover local skills on developer devices | Drop a skill folder on a managed laptop; it is found |
| Record owner, purpose, version and data classes per skill | Open any record; all four fields are filled or flagged |
| Link skills to the tools and MCP servers they call | Pick a skill that calls an MCP server; the link is shown |

## 2. Authoring standards: is every skill built the same way?

| Requirement | Acceptance test |
|---|---|
| Frontmatter validated against the format | Upload a skill with a 70-character name; it is rejected |
| Description states when the skill should trigger | Reviewer can tell from the description alone when it loads |
| No secrets in skill files | A skill containing an API key is flagged before approval |
| Scripts declare dependencies and network use | A script that calls an external URL is flagged |

## 3. Review and approval: does a person check before rollout?

| Requirement | Acceptance test |
|---|---|
| Publishing to the organisation requires review | A user publishes; it waits for approval |
| Code in skills is reviewed like any other code | Approval record shows who read the scripts |
| Risk tier set per skill | A skill touching customer data is tiered higher than a formatting skill |
| Approval tied to a version, not a name | Change one line; the new version needs approval again |

## 4. Distribution and access: who gets which skill?

| Requirement | Acceptance test |
|---|---|
| Assign skills to groups from the identity provider | Remove a user from a group; the skill disappears for them |
| Separate policies for chat apps and coding agents | Block sync to developer tools for one skill; it stays in chat only |
| Users can turn a provisioned skill off but not delete it | Attempt deletion as a user; it fails |
| Peer sharing can be limited or disabled | Turn off sharing; the share option disappears |

## 5. Runtime monitoring: what happens when skills run?

| Requirement | Acceptance test |
|---|---|
| Log when a skill is invoked, by whom and by which agent | Run a skill; an event appears with user and agent |
| Alert on unusual spread or use | Share one skill widely in a day; an alert fires |
| Watch for policy violations in skill-driven actions | A skill attempts a blocked tool call; it is blocked and logged |
| Cost of skill-driven model calls is visible | Usage of a skill shows token cost |

## 6. Change and retirement: can you update or pull a skill safely?

| Requirement | Acceptance test |
|---|---|
| Version history with diffs | Compare two versions; changes are shown line by line |
| Roll back to a previous version | Roll back; users get the old version within the stated time |
| Revoke a skill everywhere | Revoke; it stops loading in chat and on devices |
| Owner departure triggers reassignment | Deactivate an owner in the identity provider; the skill is flagged |

## How should you use this list?

1. **Rank it.** Mark each requirement must, should or could. Not every enterprise needs device discovery on day one.
2. **Sequence it.** Inventory and review first, runtime monitoring second, advanced retirement flows third.
3. **Test it.** Run the acceptance tests in the pilot, with your own skills, not vendor samples.
4. **Map gaps to platform limits.** Some gaps belong to the skill platform, not your governance tool. Note which.

For the Anthropic-specific admin controls, see [how to govern skills in Claude Team and Enterprise](/blog/how-to-govern-skills-in-claude-team-and-enterprise). For coding agents, see [how to govern Claude Code in the enterprise](/blog/govern-claude-code-in-the-enterprise). Skills often call MCP servers, so pair this list with [how to govern MCP servers](/blog/how-to-govern-mcp-servers).

## How Opencontroller does this

Opencontroller discovers and manages skills in Anthropic Enterprise and keeps them in a governed registry with agents, versions, tools and MCP servers. Opencontroller Edge, installed by IT via MDM, extends control to developer devices, coding agents, skills and local MCP servers. Policy as code applies across the estate. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**What fields does a SKILL.md file require?**
Anthropic's format requires `name` (lowercase letters, numbers and hyphens, up to 64 characters) and `description` (up to 1,024 characters) in YAML frontmatter.

**Should skill approval be tied to a version?**
Yes. Approving a name lets later edits bypass review. Approve each version.

**Which skills requirements should come first?**
Inventory and review. You cannot monitor or retire what you have not found and approved.

## Key facts

Enterprise requirements for AI agent skills management fall into six lifecycle areas: inventory (org, user-created and local device skills, owner, version, links to tools and MCP servers), authoring standards (frontmatter validation, trigger descriptions, no secrets, declared network use), review and approval (required review, code review, risk tier, version-bound approval), distribution and access (identity-provider groups, separate chat and coding-agent policies, sharing limits), runtime monitoring (invocation logs, spread alerts, policy enforcement, cost), and change and retirement (version diffs, rollback, revoke everywhere, owner-departure reassignment). Each requirement needs an acceptance test. Anthropic's SKILL.md format requires name (lowercase, numbers, hyphens, up to 64 characters) and description (up to 1,024 characters). Opencontroller by Lyzr discovers and manages skills in Anthropic Enterprise, and Opencontroller Edge extends control to skills on developer devices. Pricing starts from $250K per deployment per year.

## Related

- [How do you govern skills in Claude Team and Enterprise?](https://opencontroller.lyzr.com/blog/how-to-govern-skills-in-claude-team-and-enterprise.md)
- [What is agent skills governance?](https://opencontroller.lyzr.com/blog/what-is-agent-skills-governance.md)
- [How do you govern Claude Code across an enterprise?](https://opencontroller.lyzr.com/blog/govern-claude-code-in-the-enterprise.md)
- [More on Tools and MCP](https://opencontroller.lyzr.com/blog/cluster/tools-and-mcp)
